Getting started
This guide shows how to get started with Model Registry and run a few examples using the command line or Python clients.
At this time, the Model Registry does not include a web-based User Interface (UI), therefore this documentation focuses on backend services and APIs.
For an overview of the logical model of model registry, check the Model Registry logical model. The logical model is exposed via the Model Registry REST API.
Prerequisites
To follow along the examples in this guide, you will need a Kubeflow installation and the Model Registry installed:
- Kubeflow
- Model Registry
- Python >= 3.9
Setup
To use Model Registry on a notebook you should first install the Python client:
!pip install --pre model-registry=="0.2.3a1"
Note that depending on your environment there might be conflicting dependency versions for packages that depend on
pydantic
.
You can get a client pointing to your deployed Model Registry from the previous steps:
from model_registry import ModelRegistry
registry = ModelRegistry(
server_address="http://model-registry-service.kubeflow.svc.cluster.local",
port=8080,
author="your name",
is_secure=False
)
For more information on client setup and capabilities, refer to the Model Registry Python client documentation.
Register metadata
You can use the register_model
method to index a model’s artifacts and its metadata, for instance:
rm = registry.register_model(
"mnist",
"https://github.com/tarilabs/demo20231212/raw/main/v1.nb20231206162408/mnist.onnx",
model_format_name="onnx",
model_format_version="1",
version="v0.1",
description="lorem ipsum mnist",
metadata={
"accuracy": 3.14,
"license": "apache-2.0",
}
)
Retrieving metadata
Continuing on the previous example, you can use the following methods to retrieve the metadata associated with a given Model Artifact:
model = registry.get_registered_model("mnist")
print("Registered Model:", model, "with ID", model.id)
version = registry.get_model_version("mnist", "v0.1")
print("Model Version:", version, "with ID", version.id)
art = registry.get_model_artifact("mnist", "v0.1")
print("Model Artifact:", art, "with ID", art.id)
These can be used to create a KServe inference endpoint.
Deploy an inference endpoint
Normally you would need to provide your deployment metadata manually resulting in an error-prone process, especially when such data has to be gathered from several sources. Using Model Registry ensures simplified access to accurate metadata, and enables you to automate deployment based on the Model Registry values, as also shown in the example below.
Note: the provided example uses the Model Registry Python client and KServe Python SDK. You can analogously make use of the Model Registry REST APIs, and your own Add-on SDK as needed.
You can use the retrieved metadata from the previous step with the KServe Python SDK to create an inference endpoint, for example:
from kubernetes import client
import kserve
isvc = kserve.V1beta1InferenceService(
api_version=kserve.constants.KSERVE_GROUP + "/v1beta1",
kind=kserve.constants.KSERVE_KIND,
metadata=client.V1ObjectMeta(
name="mnist",
namespace=kserve.utils.get_default_target_namespace(),
labels={
"modelregistry/registered-model-id": model.id,
"modelregistry/model-version-id": version.id,
},
),
spec=kserve.V1beta1InferenceServiceSpec(
predictor=kserve.V1beta1PredictorSpec(
model=kserve.V1beta1ModelSpec(
storage_uri=art.uri,
model_format=kserve.V1beta1ModelFormat(
name=art.model_format_name, version=art.model_format_version
),
runtime="kserve-ovms",
protocol_version="v2",
)
)
),
)
ks_client = kserve.KServeClient()
ks_client.create(isvc)
An inference endpoint is now created, using the artifact metadata retrieved from the Model Registry (previous step), specifying the serving runtime to be used to serve the model, and references to the original entities in Model Registry.
Next steps
- Get involved:
- Model Registry working group: https://www.kubeflow.org/docs/about/community/#kubeflow-community-calendars
- https://github.com/kubeflow/model-registry
- Feedback:
Alpha
This Kubeflow component has alpha status with limited support. See the Kubeflow versioning policies. The Kubeflow team is interested in your feedbackabout the usability of the feature.Feedback
Was this page helpful?
Glad to hear it! Please tell us how we can improve.
Sorry to hear that. Please tell us how we can improve.
Last modified July 16, 2024: MR: update getting started (#3801) (1b152f5)